A hybrid machine learning approach for train trajectory reconstruction under interruptions considering passenger demand

This paper applies a hybrid data-driven prediction and optimization method to study the train trajectory reconstruction under interruption conditions. A deep reinforcement learning model, called Proximal Policy Optimization (PPO), is first used to obtain timetable rescheduling schemes, by considering the train operation constraints. Then, the dwelling times of each train at each station under interruption conditions are predicted based on a machine-learning model. Train trajectories are obtained by combining the results of the PPO model, the prediction model, and the train arrival/departure constraints. The practical case of the Wuhan to Guangzhou high-speed railway shows that (1) the PPO model is better than that obtained by other standard reinforcement learning models, with over 12.7% improvements in terms of train delays;(2) the proposed model can be trained off-line and called quickly; (3) the proposed train trajectory reconstruction method is better than the controller’s on-site decision, with approximately 20.5% reduction in train delays.